PDSXray

DSXray is a high-quality benchmark dataset dedicated to <b>X-ray object detection under physical domain shift (PDS) scenarios</b>. It aims to address the generalization performance degradation of models caused by physical factors such as device differences, environmental changes, and target morphological transformations in security inspection, medical imaging, and other fields.<br>The dataset features a carefully designed multi-source domain structure that systematically isolates endogenous domain shifts (EDS, e.g., background or texture variations) from physical domain shifts (PDS, e.g., changes in object angle, scale, occlusion, or imaging device parameters). It comprises two core subsets: a synthetic scenario subset (\texttt{PDS-EDS-Synthetic}) and a real-world scenario subset (\texttt{PDS-Aggregated}). The real-world subset integrates authentic security X-ray images from sources like KDXray, CLCXray, and PIDray, covering over 10 high-priority target categories such as knives, lighters, power banks, and liquid containers, with annotations including precise bounding boxes and class labels to support fine-grained performance analysis.<br>PDSXray not only provides a standardized evaluation benchmark for quantifying the impact of domain shifts but also reveals the sensitivity of different models and object categories to physical domain shifts through multi-dimensional baseline experiments (e.g., performance comparisons of YOLO-series models). It serves as a reliable tool for developing and validating domain adaptation and robust detection algorithms, ultimately enhancing the practical value of X-ray object detection in complex real-world scenarios.